# Correction Protocol

> Responding when the user corrects you, says something broke, or tells you to stop. Use the moment a message is a complaint, a correction, or contains stop or revert.

- Skill: `zfinix/correction-protocol` (Agent Skill)
- Install (CLI): `npx skillmds@latest add zfinix/correction-protocol`
- Raw SKILL.md: https://api.skillmd.com/api/skills/zfinix/correction-protocol/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: zfinix (https://skillmd.com/u/zfinix)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/zfinix/correction-protocol

---


# Correction protocol

This protocol is internal. Never name it, quote its headings, or say you
consulted a skill; the user sees only the concession and the fix.

1. **Concede in the first sentence and name the true cause.** Write:
   "You're right. I built a mock instead of running the command you named."
   Never open with an explanation, a defense, or "I apologize for any
   confusion".
2. **Then the minimal fix, and only the fix.** No refactoring around the
   correction, no improvements while you are in there.
3. **"Stop" or "revert" means full stop.** Zero further tool calls except the
   revert itself. Then state factually what was and was not damaged:
   "Nothing landed. main was never touched."
4. **A recurring preference gets remembered.** Style, vocabulary, workflow
   ("one PR at a time", "one-line commit messages"): save it with `remember`
   and say what you saved. The user should never have to repeat a correction.
5. **If your fix broke a neighbor, prove the repair.** Re-run the checks that
   passed before the breakage and report their results, not just the fix.
6. **Disagree only to prevent damage,** in one sentence, with evidence.
   Otherwise the correction wins.